A corrugated board warping deformation self-tuning PID control system based on a fuzzy neural network

Abstract

This invention relates to a fuzzy neural network-based self-tuning PID control system for corrugated board warpage deformation, applied in the field of corrugated board production technology. The system includes: a data acquisition unit for real-time acquisition of raw paper moisture content, preheating cylinder temperature, ambient relative humidity, production line speed, and composite tension; a warpage detection device for detecting the amount of board warpage; a fuzzy neural network controller for outputting PID parameter adjustment values; a PID parameter self-tuning unit for real-time calculation of PID parameters; a PID controller for calculating control values; an actuator unit for adjusting operating parameters; and a graded trigger optimization module for graded online fine-tuning based on a superior product threshold and a qualified product threshold: no update when deviation ≤ superior product threshold, data is stored only when superior product threshold < deviation ≤ qualified product threshold, and an update is triggered when deviation > qualified product threshold. This invention solves the problems of low control accuracy and frequent ineffective adjustments in existing technologies, improving system stability and adaptability.

Description

Technical Field

[0001] This invention relates to the field of corrugated board production technology, and more specifically, to a self-tuning PID control system for corrugated board warping deformation based on a fuzzy neural network. Background Technology

[0002] Corrugated cardboard is widely used in the packaging industry due to its advantages such as light weight, high strength, and good cushioning performance. During the production of corrugated cardboard, warping and deformation frequently occur due to the combined effects of various factors, including the moisture content of the base paper, the temperature of the preheating cylinder, ambient temperature and humidity, production line speed, and lamination tension. Warping not only affects the appearance quality of the cardboard but also leads to a decrease in the processing precision of subsequent processes such as printing, slotting, and die-cutting, and can even result in defective products.

[0003] Currently, the control methods for corrugated cardboard warpage mainly rely on manual adjustment based on operator experience, or on closed-loop control of individual parameters (such as preheating cylinder temperature) using conventional PID controllers. However, warpage is a complex process characterized by multi-factor coupling, nonlinearity, and large time lag. Conventional PID control struggles to adapt to changing operating conditions, faces difficulties in parameter tuning, and results in unsatisfactory control performance.

[0004] In recent years, intelligent control methods such as fuzzy control and neural networks have been introduced into the field of industrial process control. For example, existing technologies have combined neural networks and fuzzy control for motor synchronization control in corrugated cardboard production lines to indirectly influence warping; other studies have disclosed online fine-tuning mechanisms for fuzzy neural network PID controllers. However, these methods either fail to directly address the specific physical mechanisms of corrugated cardboard warping by selecting parameters, resulting in limited control accuracy; or, although online learning mechanisms are introduced, they fail to consider the frequent ineffective adjustments caused by sensor noise, transient interference, and other factors in actual production, affecting the stability and lifespan of the control system.

[0005] Therefore, there is an urgent need for an adaptive control system that can address the multi-field coupling characteristics of corrugated cardboard warping while also ensuring control accuracy and online learning stability. Summary of the Invention

[0006] The present invention aims to solve the technical problems of low warpage control accuracy and poor adaptability of corrugated cardboard in the prior art, while improving the stability of online learning.

[0007] To achieve the above objectives, the present invention provides a self-tuning PID control system for the warpage deformation of corrugated cardboard based on a fuzzy neural network, comprising:

[0008] The data acquisition unit includes multiple sensors for real-time acquisition of process parameters during the corrugated cardboard production process. These process parameters include at least the moisture content of the base paper, the temperature of the preheating cylinder, the relative humidity of the environment, the speed of the production line, and the composite tension.

[0009] The warpage detection device is installed at the end of the corrugated cardboard production line to detect the warpage direction and amount of the cardboard in real time and generate the actual warpage value.

[0010] The fuzzy neural network controller is connected to both the data acquisition unit and the warpage detection device. It incorporates a built-in, offline-trained artificial neural network model. Taking process parameters as input, it outputs the adjustment amount of the PID control parameters, including the proportional coefficient adjustment. Integral coefficient adjustment and differential coefficient adjustment amount ;

[0011] The PID parameter self-tuning unit is connected to the fuzzy neural network controller and is used to calculate the PID control parameters at the current moment in real time based on the reference PID parameters and the adjustment amount.

[0012] A PID controller, connected to a PID parameter self-tuning unit, is used to calculate the control quantity based on the current PID control parameters and warpage deviation value.

[0013] The actuator unit, connected to the PID controller, is used to receive control inputs and adjust the operating parameters of the corrugated cardboard production line.

[0014] The graded triggering optimization module is connected to the warpage detection device and the fuzzy neural network controller, respectively, to monitor the actual warpage value in real time and determine the superior grade threshold as defined in the corrugated cardboard product quality grade standard. and qualified product threshold Hierarchical trigger control strategy ( ):

[0015] When the actual warpage deviation satisfy At the same time, keep the current weights of the artificial neural network model unchanged;

[0016] when At that time, the process parameters and deviation data at the current moment are stored in the historical database, but the online weight update of the artificial neural network model is not triggered.

[0017] when At that time, the fuzzy neural network controller is triggered to perform online fine-tuning and update the weights of the artificial neural network model;

[0018] Among them, the threshold for superior grade products and qualified product threshold It is pre-set according to the warpage limits for superior and qualified products specified in the national standards for the target corrugated cardboard products.

[0019] As a further improvement to the present invention, the process parameters are selected based on the multi-field coupling physical mechanism of corrugated cardboard warping deformation:

[0020] The moisture content of the base paper and the relative humidity of the environment together constitute the factors affecting the moisture field, which are used to characterize the influence of the moisture migration trend of the paperboard on warping during the production process.

[0021] The temperature of the preheating cylinder constitutes a factor affecting the temperature field and is used to characterize the effect of the degree of heat absorption of the cardboard in the preheating zone on warping.

[0022] Production line speed constitutes a time-space coupled influencing factor, used to characterize the impact of the dwell time of the cardboard in each process section on moisture evaporation and heat transfer sufficiency;

[0023] Composite tension constitutes a mechanical field influencing factor, used to characterize the effect of tensile stress on the stress distribution within the paperboard when the face paper and corrugated paper are bonded together.

[0024] As a further improvement of the present invention, the fuzzy neural network controller includes a fuzzification processing module and a neural network training module:

[0025] The blurring processing module includes:

[0026] Membership function definition unit, used to define multiple fuzzy subsets for each process parameter, and set membership functions for each fuzzy subset;

[0027] The fuzzy rule base storage unit is used to store the fuzzy control rule base built based on the multi-field coupling mechanism;

[0028] The fuzzy inference unit is used to perform fuzzy inference based on the continuous values ​​of each process parameter at the current time, combined with the membership function and fuzzy control rules, to generate a fuzzy input vector.

[0029] The fuzzy rule base should include at least:

[0030] Rule R1: IF Preheat cylinder temperature = high AND paper moisture content = high AND machine speed = fast THEN Warping tendency = severe upward warping;

[0031] Rule R2: IF Preheat cylinder temperature = low AND base paper moisture content = low AND machine speed = slow THEN Warp tendency = slight downward warping;

[0032] Rule R3: IF Composite tension = High AND Preheat cylinder temperature = Normal AND Base paper moisture content = Normal THEN Warp tendency = Lateral warp;

[0033] Rule R4: IF Ambient relative humidity = high AND base paper moisture content = normal AND preheating cylinder temperature = normal THEN Warping tendency = slight downward warping;

[0034] Rule R5: IF Ambient relative humidity = low AND base paper moisture content = normal AND preheating cylinder temperature = normal THEN Warping tendency = slight upward warping;

[0035] The neural network training module has a built-in artificial neural network model that takes a fuzzy input vector as input and outputs the adjustment amount of the PID control parameters.

[0036] As a further improvement of the present invention, in the graded triggering optimization module, the triggering condition for online fine-tuning is: continuous Within each sampling period, the absolute value of the deviation between the actual warp value and the target warp value exceeded the qualified product threshold. ,in, It is an integer between 3 and 10;

[0037] Online fine-tuning employs incremental learning algorithms, including stochastic gradient descent or mini-batch gradient descent, to update the weights of the artificial neural network model.

[0038] The update step size is set to the initial learning rate during the offline training phase. to .

[0039] As a further improvement of the present invention, the artificial neural network model in the fuzzy neural network controller adopts a deep belief network or convolutional neural network structure to extract the deep nonlinear coupling features between process parameters and warping deformation; the deep belief network or convolutional neural network is trained offline by combining unsupervised pre-training and supervised fine-tuning.

[0040] As a further improvement of the present invention, in the graded triggering optimization module, the superior product threshold is... and qualified product threshold It is a variable threshold that is dynamically adjusted based on warping data under the same or similar production conditions stored in a historical database, through cluster analysis or regression prediction models.

[0041] As a further improvement of the present invention, a baseline PID parameter adaptive initialization module is also included. This module is connected to a historical database and is used to retrieve the optimal baseline PID parameter under similar working conditions from the historical database before each production start, based on the cardboard specifications and raw material batch information of the current production order, and use it as the initial baseline parameter. , , .

[0042] As a further improvement of the present invention, the graded triggering optimization module also includes a working condition similarity calculation unit, which is used to calculate the working condition similarity between the current production batch and the historical batches stored in the historical database. The working condition similarity is calculated based on at least one factor among the raw paper batch, the rate of change of ambient temperature and humidity, and the equipment running time. When the working condition similarity exceeds a preset threshold, the historical threshold under the same or similar working conditions is preferentially used as the initial superior product threshold for the current batch. and qualified product threshold .

[0043] As a further improvement of the present invention, a control quantity feedforward compensation module is also included. This module is connected to the data acquisition unit and the PID controller, and is used to pre-calculate the warping trend change based on the current rate of change in the moisture content of the raw paper and the rate of change in the production line speed, and generate a feedforward compensation quantity that is superimposed on the control quantity output by the PID controller. The formula for calculating the feedforward compensation quantity is: ,in, This refers to the moisture content of the base paper. For production line speed, and The compensation coefficient is determined based on historical data.

[0044] As a further improvement of the present invention, the artificial neural network model in the fuzzy neural network controller is trained offline using a multi-objective optimization algorithm. The multi-objective includes at least two of the following: minimizing warpage deviation, minimizing the rate of change of PID parameter adjustment, and minimizing the number of actuator adjustments. The multi-objective optimization algorithm uses the NSGA-II or MOPSO algorithm to generate a Pareto optimal solution set, and selects suitable neural network weights from the Pareto optimal solution set as initial weights according to the current production conditions.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention selects five key process parameters—base paper moisture content, preheating cylinder temperature, ambient relative humidity, production line speed, and composite tension—and categorizes them into moisture field, temperature field, mechanical field, and spatiotemporal coupling field, respectively. This comprehensively covers the main physical factors affecting the warping of corrugated cardboard, providing accurate input features for subsequent intelligent control and significantly improving the accuracy of warping control.

[0047] 2. This invention sets a threshold for superior quality products. and qualified product threshold It adopts different strategies (hold, store only, trigger fine-tuning) according to the actual deviation range, which solves the problem of frequent ineffective adjustment caused by instantaneous interference or small fluctuations in traditional online learning. While ensuring the rate of excellent products, it effectively extends the model stabilization period and improves the stability of the control system.

[0048] 3. This invention links the threshold setting with the national standard's superior / qualified product limits, making the control target directly correspond to the product quality level, thus improving the system's industry adaptability and practical application value.

[0049] 4. The five dedicated fuzzy rules constructed based on the multi-field coupling mechanism in this invention accurately reflect the causal relationship between the changes in each parameter and the warping trend, making fuzzy inference more targeted and further improving control accuracy.

[0050] 5. The synergistic effect of mechanisms such as continuous N-cycle triggering conditions, small step size updates, and feedforward compensation effectively suppresses system oscillations and improves the robustness of the control process to noise and interference.

[0051] 6. The introduction of modules such as dynamic threshold adjustment, adaptive initialization of baseline parameters, and similarity calculation of operating conditions enables the system to adapt to different production batches and environmental changes, achieving true adaptive intelligent control. Attached Figure Description

[0052] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0054] Example 1:

[0055] Figure 1 This paper presents a self-tuning PID control system for the warpage deformation of corrugated cardboard based on a fuzzy neural network, comprising the following components:

[0056] 1. Data Acquisition Unit

[0057] The data acquisition unit includes multiple sensors for real-time acquisition of process parameters during corrugated board production. These parameters include at least the moisture content of the base paper, the temperature of the preheating cylinder, the relative humidity of the environment, the production line speed, and the laminating tension. Multiple sensors are installed at key locations on the production line, for example:

[0058] Paper moisture content sensor: Installed at the paper rack, using an infrared moisture meter or microwave moisture meter, it measures the initial moisture content of the face paper and corrugated paper entering the production line in real time, with a measurement accuracy of ±0.1%.

[0059] Preheating cylinder temperature sensor: Installed on the surface or inside the preheating cylinder, using a thermocouple or resistance temperature detector, to measure the working temperature of the preheating cylinder in real time, with a measurement range of 80℃-200℃ and an accuracy of ±0.5℃.

[0060] Ambient temperature and humidity sensor: Installed in the production line workshop, away from heat sources and ventilation openings, to measure the relative humidity (0-100%RH) and temperature (-20℃-80℃) of the ambient air in real time, with an accuracy of ±2%RH and ±0.3℃.

[0061] Production line speed sensor: A rotary encoder is installed on the main drive roller to measure the running line speed of the cardboard in real time. The measurement range is 0-300m / min and the accuracy is ±0.1m / min.

[0062] Composite tension sensor: Installed at the bearing seat of tension roller or floating roller, it adopts strain gauge or piezoelectric tension sensor to measure the tension of face paper and corrugated paper in real time. The measurement range is 0-500N / m and the accuracy is ±1N / m.

[0063] The selection of the above five process parameters is based on the multi-field coupled physical mechanism of corrugated cardboard warpage: the moisture content of the base paper and the relative humidity of the environment together constitute the moisture field influencing factor, used to characterize the impact of the moisture migration trend of the cardboard during production on warpage; the preheating cylinder temperature constitutes the temperature field influencing factor, used to characterize the impact of the degree of heat absorption of the cardboard in the preheating zone on warpage; the production line speed constitutes the time-space coupled influencing factor, used to characterize the impact of the residence time of the cardboard in each process section on the sufficiency of moisture evaporation and heat transfer; and the composite tension constitutes the mechanical field influencing factor, used to characterize the impact of the tensile stress when the face paper and corrugated paper are bonded on the stress distribution within the cardboard. These five parameters comprehensively cover the main physical fields affecting warpage, providing accurate input characteristics for subsequent intelligent control.

[0064] All sensors are equipped with signal conditioning circuits and A / D conversion modules, with sampling frequencies set from 10Hz to 50Hz (adjustable according to production line speed). The collected data is transmitted to the central controller via fieldbus (such as PROFIBUS, Modbus) or industrial Ethernet.

[0065] 2. Warpage detection device

[0066] The warpage detection device is installed at the end of the corrugated cardboard production line, before the cardboard is stacked. This embodiment uses a laser displacement sensor array, consisting of 5-10 laser displacement sensors arranged at equal intervals along the width of the cardboard. Each sensor has a measurement range of ±50mm and an accuracy of ±0.1mm. The sensor array scans the surface contour of the cardboard in real time, generating a warpage curve along the width direction.

[0067] The detection device also includes a data processing unit for:

[0068] Filter the multi-point measurement data to eliminate random noise;

[0069] The least squares method is used to fit the warped curve and calculate the maximum warping (i.e., the vertical distance between the highest and lowest points of the curve).

[0070] Determine the direction of warping (upward warping, downward warping, or horizontal warping) based on the relative positions of the highest and lowest points.

[0071] The output of the warpage detection device is the actual warpage value. (Unit: mm) and warping direction indicator; sampling period synchronized with data acquisition unit.

[0072] 3. Fuzzy Neural Network Controller

[0073] The fuzzy neural network controller is the core intelligent control unit of this invention. It is implemented by an industrial computer or embedded system and runs a fuzzy neural network algorithm internally.

[0074] The controller receives process parameters from the data acquisition unit via Ethernet. (Corresponding to the moisture content of the base paper, the temperature of the preheating cylinder, the relative humidity of the environment, the speed of the production line, and the composite tension, respectively) and the actual warp value of the warp detection device. It also outputs the adjustment amount of the PID parameters. .

[0075] 4. PID parameter self-tuning unit

[0076] The PID parameter self-tuning unit is a software module that runs within the central controller. Its function is to tune the PID parameters based on a reference PID parameter. Adjustment amount of the output of the fuzzy neural network controller Calculate the PID control parameters in real time:

[0077]

[0078]

[0079]

[0080] The initial values ​​of the baseline PID parameters can be determined in one of the following ways:

[0081] The Ziegler-Nichols tuning method was used to obtain the result.

[0082] Input is made by the operator based on experience;

[0083] The baseline PID parameter adaptive initialization module retrieves the parameters from the historical database (see below).

[0084] 5. PID controller

[0085] The PID controller employs a positional PID control algorithm, based on the current PID control parameters and the warpage deviation value. (in, (The target warp value is usually set to 0) Calculate the control quantity. :

[0086]

[0087] in, The sampling period is set to 0.1s-1s in this embodiment (adjusted according to the production line speed).

[0088] 6. Execution unit

[0089] The actuator unit receives the control input from the PID controller. This is then converted into specific physical adjustment actions. In this embodiment, the actuator includes:

[0090] Preheating cylinder heating valve: adopts an electric regulating valve, receives a 4-20mA current signal, controls the heating steam flow, and the adjustment range is 0-100% opening;

[0091] Tension roller drive motor: A variable frequency motor is used, which receives a 0-10V voltage signal to adjust the speed of the tension roller, thereby changing the composite tension;

[0092] Main drive variable frequency motor: The variable frequency motor receives 0-10V voltage signals to adjust the speed of the main drive rollers of the production line, thereby changing the speed of the production line.

[0093] The actuator unit is also equipped with a driver and a servo amplifier to ensure accurate conversion and execution of control signals.

[0094] 7. Hierarchical Trigger Optimization Module

[0095] The tiered triggering optimization module is the core innovative module of this invention, running as software within the central controller. This module monitors the actual warpage value in real time. Calculate the deviation And according to the preset superior product threshold and qualified product threshold ( ) Execute the hierarchical trigger control strategy.

[0096] In this embodiment, the threshold for superior grade products and qualified product threshold The warpage limits are set according to the national standard GB / T 6544-2008 "Corrugated Board" for superior and qualified products. Specifically, for Class A corrugated board, the standard stipulates that the warpage of superior products is ≤3mm, and the warpage of qualified products is ≤5mm. Therefore, the following limits are set: , For other types of cardboard, adjustments can be made according to the corresponding limits in the standard.

[0097] The specific execution flow of the tiered trigger control strategy is as follows:

[0098] when At this time, the cardboard quality meets the superior standard, the control system considers the current operating condition to be good, and the weights of the artificial neural network model remain unchanged. At this point, only conventional PID control is performed, without triggering any learning or update operations.

[0099] when At this time: the cardboard quality is within the acceptable range but has not reached the superior grade standard. At this point, the system will display the current process parameters. and deviation data The data is stored in a historical database for subsequent offline training and data analysis, but does not trigger online weight updates for the artificial neural network model. This design aims to avoid frequent online learning due to minor fluctuations (such as sensor noise or transient interference), thereby maintaining the stability of the control system.

[0100] when When the cardboard quality exceeds the acceptable standard, resulting in severe warping and deformation, the system immediately triggers online fine-tuning. It uses an incremental learning algorithm to update the weights of the artificial neural network model, adapting the model to the new operating conditions and outputting a more suitable PID adjustment, thereby quickly suppressing warping.

[0101] 8. Historical Database

[0102] The historical database uses a SQL Server or MySQL relational database and is deployed on the solid-state drive of the central controller. The database records the following information:

[0103] Timestamp (accurate to milliseconds)

[0104] Original values ​​of 5 process parameters

[0105] Actual warpage value and deviation

[0106] Weight snapshot of an artificial neural network model (optional, saved as needed).

[0107] Historical records of PID control parameters

[0108] Production order information (cardboard specifications, raw material batches, etc.)

[0109] Environmental information (season, weather, etc., can be entered by the operator).

[0110] The capacity of the historical database is set according to the production scale. In this embodiment, the production data of the most recent 3 months is retained for offline training, data analysis, dynamic adjustment of thresholds and initialization of benchmark parameters.

[0111] Example 2:

[0112] This embodiment illustrates the specific implementation of a fuzzy neural network controller.

[0113] 2.1 Blurring Processing Module

[0114] The fuzzification module converts five continuous process parameters into fuzzy input vectors. Each process parameter defines five fuzzy subsets: {Very Low (VL), Low (L), Normal (M), High (H), Very High (VH)}. The membership function uses a trigonometric function; taking the moisture content of the base paper as an example:

[0115]

[0116]

[0117]

[0118]

[0119] Among them, parameters The parameters are determined based on the physical range of the process parameters and expert experience. For example, the typical range for the moisture content of the base paper is 4%-10%, and can be set as follows:

[0120] VL: 4%-5.2% (parameter) )

[0121] L: 5%-6.4% (parameter) )

[0122] M: 6%-7.6% (parameter) )

[0123] H: 7.2%-8.8% (parameter) )

[0124] VH: 8.5%-10% (parameter) )

[0125] The definitions of fuzzy subsets for other parameters are similar and will not be repeated here.

[0126] 2.2 Fuzzy Rule Base

[0127] The fuzzy rule base stores fuzzy control rules constructed based on multi-field coupling mechanisms. The fuzzy rule base constructed in this embodiment includes at least the following five core rules:

[0128] Rule Number Prerequisites in conclusion R1 Preheating cylinder temperature = high AND paper moisture content = high AND machine speed = fast Warping trend = severe upward warping R2 Preheating cylinder temperature = low AND paper moisture content = low AND machine speed = slow Warping tendency = slight downward warping R3 Composite tension = High AND Preheating cylinder temperature = Normal AND Base paper moisture content = Normal Warping tendency = lateral warping R4 Ambient relative humidity = high AND paper moisture content = normal AND preheating cylinder temperature = normal Warping tendency = slight downward warping R5 Ambient relative humidity = low AND paper moisture content = normal AND preheating cylinder temperature = normal Warp tendency = slight upward warping

[0129] The confidence factor for each rule is set to 1.0, and can be fine-tuned based on the actual application effect.

[0130] 2.3 Fuzzy Inference Unit

[0131] The fuzzy inference unit employs the Mamdani inference method. For each sampling time, based on the current values ​​of the five process parameters, the premise satisfaction degree of each rule is calculated (taking the minimum membership degree of each premise). Then, a weighted average method is used for defuzzification to obtain the fuzzy input vector. ,in, The total number of fuzzy subsets (in this embodiment, it is...) ).

[0132] 2.4 Basic Implementation of Neural Network Training Module (Based on BP Network)

[0133] To facilitate understanding of the core control logic of this invention, this embodiment first uses the classic BP neural network as an example to illustrate the basic structure and working principle of the neural network training module. The BP network structure is as follows:

[0134] Input layer: The number of nodes is equal to the dimension of the fuzzy input vector, i.e., 25 nodes.

[0135] Hidden layer: In this embodiment, it is set to 12 nodes, and the activation function is the Sigmoid function.

[0136] Output layer: 3 nodes, corresponding to... The activation function is a linear function.

[0137] The offline training of the BP network uses the Levenberg-Marquardt algorithm, with training samples drawn from historical production data (including process parameters under different operating conditions and their corresponding optimal PID parameter adjustments). The target error is set to 0.001. For online fine-tuning, stochastic gradient descent (SGD) is used, with the update step size set to 1 / 8 of the offline learning rate.

[0138] 2.5 Preferred Implementation Scheme (Based on Deep Belief Network DBN)

[0139] To further enhance the network's ability to extract multi-field coupling features, this invention preferably employs a Deep Belief Network (DBN) structure. A DBN is composed of multiple Restricted Boltzmann Machines (RBMs) stacked together, and its structure is as follows:

[0140] Input layer: 25 nodes (corresponding to fuzzy input vectors)

[0141] First hidden layer: 20 nodes (RBM1)

[0142] Second hidden layer: 15 nodes (RBM2)

[0143] Output layer: 3 nodes, corresponding to... The activation function is a linear function.

[0144] The training is divided into two phases:

[0145] Unsupervised pre-training: Each RBM is trained layer by layer using the contrastive divergence (CD-1) algorithm with a learning rate of 0.01 and 100 iterations per layer. The learning rate and number of iterations are determined through cross-validation: 80% of the historical data is randomly selected as the training set and 20% as the validation set to test the reconstruction error under different parameter combinations. The parameter combination with the smallest error on the validation set is then selected. After pre-training, the initial weights for each layer are obtained.

[0146] Supervised fine-tuning: Based on the pre-trained weights, the backpropagation (BP) algorithm is used to globally fine-tune the entire network. The objective function is the sum of squared warp deviations, the learning rate is set to 0.001, and the iterations are 500. Cross-validation is also used in the fine-tuning phase to prevent overfitting. Training is terminated early when the validation set error does not decrease for 10 consecutive iterations.

[0147] DBN can effectively extract deep nonlinear coupling features between process parameters and warpage deformation, and is particularly suitable for the multi-field coupling scenario of this invention. During online fine-tuning, to avoid destroying the pre-trained features, only the last two layers can be fine-tuned, with the fine-tuning step size set to 1 / 10 of the offline learning rate.

[0148] 2.6 Another preferred implementation scheme (based on Convolutional Neural Network CNN)

[0149] As another preferred embodiment, the present invention can also employ a convolutional neural network (CNN) structure. Input data construction: A 5×10 two-dimensional feature map composed of the process parameters at the current time and the previous 9 sampling times is used as input, with a sampling interval of 0.1s. Specifically, for the... Each sampling time, take that time and... The five process parameter values ​​at ten different time points form a 5x10 matrix. Each row corresponds to a time series of a process parameter. To eliminate the influence of dimensions, the time series of each process parameter is normalized to the interval [0,1]. The normalization formula is as follows: ,in, These are the historical minimum and maximum values ​​of this parameter. The CNN structure is as follows:

[0150] Convolutional layer C1: 32 3×3 convolutional kernels, stride 1, activation function ReLU

[0151] Pooling layer P1: 2×2 max pooling, step size 2

[0152] Convolutional layer C2: 64 3×3 convolutional kernels, stride 1, activation function ReLU

[0153] Pooling layer P2: 2×2 max pooling, step size 2

[0154] Fully connected layer F1: 128 nodes, activation function ReLU

[0155] Output layer: 3 nodes, linear activation function

[0156] The training algorithm uses the Adam optimizer with a learning rate of 0.001, 200 iterations, and a batch size of 32. CNNs are suitable for extracting local correlations and temporal features between parameters, and can capture the impact of dynamic factors such as changes in production line speed on warping. During online fine-tuning, the parameters of the convolutional layers are fixed, and only the fully connected layers are fine-tuned to reduce computational cost.

[0157] 2.7 Online Fine-tuning Algorithm

[0158] Regardless of the network architecture used, online fine-tuning employs an incremental learning algorithm. When the hierarchical trigger optimization module determines that fine-tuning is needed, it updates the network weights in one step using stochastic gradient descent (SGD) or mini-batch gradient descent, taking the current input and output data as samples. The update step size is uniformly set to 1 / 10 to 1 / 5 of the initial learning rate during offline training to avoid system oscillations caused by sudden weight changes. For DBN and CNN, fine-tuning can be performed on only certain layers (such as the last two layers of DBN or the fully connected layers of CNN) to balance learning speed and stability, depending on the network characteristics.

[0159] To reduce the impact of online fine-tuning on the stability of the control system, the following protection mechanism is set up:

[0160] Maximum single weight change limit: Calculate the L2 norm of the current layer weight matrix. The ratio of the updated weight norm to the original norm is limited to no more than 1.1, i.e. If the limit is exceeded, the update step size will be reduced proportionally.

[0161] The minimum time interval between two consecutive fine-tuning adjustments is 10 sampling periods.

[0162] If the deviation does not improve after 5 consecutive fine-tuning attempts, pause the fine-tuning and issue an alarm.

[0163] Example 3:

[0164] This example illustrates the threshold for superior quality products. and qualified product threshold The dynamic adjustment method, and the calculation and application of working condition similarity.

[0165] 1. Dynamic threshold adjustment based on cluster analysis

[0166] The historical database stores process parameters, warpage data, and corresponding optimal thresholds for different production batches. This embodiment uses the K-means clustering algorithm to classify historical operating conditions, with each cluster center representing a typical operating condition.

[0167] For the current production batch, the system first calculates its distance to each cluster center to determine its operating condition category. Then, it extracts warpage data from all historical batches within that category and uses a multiple linear regression model to establish a mapping relationship between process parameters and warpage.

[0168]

[0169] in, For warpage, This refers to the moisture content of the base paper. To preheat the cylinder temperature, For ambient relative humidity, For vehicle speed, For composite tension, For regression coefficients, This represents the error term. The regression coefficients are obtained by fitting historical data using the least squares method.

[0170] Based on this mapping relationship, it is predicted that under the current working conditions, to achieve the superior product standard for warpage ( The required combination of process parameters, and the corresponding... and Threshold. The specific reverse calculation method is as follows: substitute the process parameter values ​​of the current operating conditions into the regression model, calculate the predicted range of warpage within the 95% confidence interval, and take the upper limit of this range as the threshold for superior products. Take 1.5 times the upper limit as the threshold for qualified products. .

[0171] The trigger conditions for dynamic adjustment are: when the specifications of the cardboard or the batch of raw materials in the production order change, or when there are significant changes in the ambient temperature and humidity (such as a day-night temperature difference exceeding 10°C or seasonal changes).

[0172] 2. Operating Condition Similarity Calculation Unit

[0173] The operating condition similarity calculation unit is used to evaluate the similarity between the current production batch and historical batches, providing a basis for threshold initialization. Similarity calculation is based on the following factors:

[0174] Raw paper batches: Raw paper batches with the same batch number are considered completely similar (similarity 1.0). For different batches, the similarity is calculated based on information such as the raw paper manufacturer, model, and basis weight (0.5-0.9).

[0175] Environmental temperature and humidity change rate: Calculate the Euclidean distance between the current environmental temperature and humidity and the historical batch temperature and humidity, and convert it into a similarity score (0.6-1.0).

[0176] Equipment runtime: The closer the equipment runtime, the higher the similarity (0.7-1.0).

[0177] Overall similarity Calculated using the weighted average method:

[0178]

[0179] Among them, weight It can be adjusted according to actual application.

[0180] when When the operating conditions are considered highly similar, the system prioritizes using historical thresholds from the same or similar operating conditions as the initial threshold for the current batch. and Otherwise, the default threshold specified by the national standard shall be used.

[0181] Example 4:

[0182] This embodiment illustrates the implementation of the baseline PID parameter adaptive initialization module.

[0183] Before each production start (such as order change, paper change, shift change), the baseline PID parameter adaptive initialization module performs the following steps:

[0184] Feature extraction: Extract key features from the current production order, including cardboard type (A-flute, B-flute, C-flute, etc.), cardboard basis weight (combination of basis weight of face paper / corrugated paper), raw material batch number, etc.

[0185] Similar working conditions retrieval: Retrieves the most similar condition to the current feature from the historical database. One historical production batch (in this embodiment, we take...) Similarity was calculated using Euclidean distance, and the feature vector dimension was 5 (cardboard type code, face paper weight, corrugated paper weight, ambient temperature, and ambient humidity).

[0186] Optimal parameter extraction: from the retrieved Extract the baseline PID parameters used in the stable production phase for each batch from the historical batches. and the corresponding average warpage .

[0187] Weighted average calculation: The weighted average baseline parameter is calculated using the reciprocal of the average warp as the weight.

[0188]

[0189] in, It is a small constant (taken as 0.1) to avoid division by zero.

[0190] Parameter output: The calculated result The output is sent to the PID parameter self-tuning unit as the initial reference parameters for this batch of production.

[0191] Example 5: Control Quantity Feedforward Compensation Module

[0192] This embodiment illustrates the implementation of the control quantity feedforward compensation module.

[0193] The feedforward compensation module is designed to address the large hysteresis characteristics of corrugated cardboard warping control, aiming to compensate in advance for warping trends caused by changes in the moisture content of the raw paper and changes in machine speed.

[0194] 1. Feedforward compensation principle

[0195] The basic idea of ​​feedforward compensation is: when the moisture content of the base paper is... Or production line speed When rapid changes occur, the warping trend will change in advance. By monitoring these rates of change, compensation can be applied in advance to counteract the impending warping.

[0196] 2. Calculation of feedforward compensation

[0197] Feedforward compensation The calculation formula is:

[0198]

[0199] in:

[0200] The rate of change in the moisture content of the base paper is calculated by the difference between consecutive sampled values:

[0201] The rate of change of production line speed is calculated using the same method as above.

[0202] and The compensation coefficient is calibrated based on historical data.

[0203] 3. Compensation coefficient calibration

[0204] Compensation coefficient and The following calibration experimental steps were used to determine:

[0205] Step 1: Establish stable operating conditions. Select a stable operating period for the production line, ensuring constant machine speed (fluctuation <1%), stable paper moisture content (fluctuation <0.1%), and warpage close to 0 (<0.5mm). Record the baseline values ​​of each parameter during this period.

[0206] Step 2: Moisture Content Step Experiment. By adjusting the temperature of the preheating cylinder or the water spray device, an approximate step change in the moisture content increment of the base paper is created. (For example, if the moisture content increases rapidly from 6% to 7%, the rate of change is approximately 0.5% / s), continuously record the moisture content change curve. and warpage change curve The sampling frequency is 10Hz and the recording time is 60s.

[0207] Step 3: Data Processing. This involves processing the records... and Low-pass filtering (cutoff frequency 1Hz) is applied to eliminate high-frequency noise. The rate of change in moisture content is calculated. And observe the delay time of the warp response. (The time difference between when the moisture content begins to change and when the warpage begins to change).

[0208] Step 4: System Identification. A first-order inertia plus delay model is used to fit the transfer function from moisture content change to warping change: .

[0209] The observations obtained in step 3 As a known quantity, the gain is obtained by fitting the experimental data using the least squares method. and time constant .

[0210] Step 5: Calculation of Compensation Coefficient. According to feedforward control theory, to offset the effect of moisture content changes on warpage, the feedforward compensation coefficient is calculated. It should be taken as the ratio of system gain to time constant, i.e. .

[0211] Step 6: Repeat the verification. Repeat steps 2-4 a total of 5 times, and take the results. The average value is used as the final calibration value.

[0212] Step 7: Vehicle Speed ​​Variation Experiment. Similarly, a step change in vehicle speed is created by changing the inverter frequency. (For example, when increasing from 100m / min to 110m / min, the rate of change is approximately 2 (m / min) / s), continuously record the vehicle speed change curve. and warpage change curve The sampling frequency was 10Hz, and the recording duration was 60s. After low-pass filtering (cutoff frequency 1Hz) of the recorded data, a first-order inertial model was used to fit the transfer function from the change in vehicle speed to the change in warpage.

[0213]

[0214] The gain was obtained by fitting the experimental data using the least squares method. According to feedforward control theory, to counteract the effect of vehicle speed changes on warpage, the feedforward compensation coefficient... It should be taken as the system gain, i.e. .

[0215] The experiment was repeated 5 times, and the average value of β was taken as the final calibration value.

[0216] Step 8: Regular recalibration. It is recommended to recalibrate every quarter or after changing raw material batches to adapt to equipment aging and changes in raw material properties.

[0217] 4. Superposition of feedforward and PID

[0218] Feedforward compensation The control quantity output by the PID controller The signals are superimposed to form the final control signal for the actuator:

[0219]

[0220] The superimposed control quantity is then limited (to prevent it from exceeding the allowable range of the actuator) before being output to the actuator unit.

[0221] Example 6: Offline Training for Multi-Objective Optimization

[0222] This embodiment illustrates a multi-objective optimization offline training method for artificial neural network models.

[0223] Traditional neural network training typically focuses on minimizing warpage deviation as a single objective, which can lead to problems such as overly drastic PID parameter adjustments and frequent actuator movements. This embodiment introduces multi-objective optimization, considering the following three objectives simultaneously:

[0224] Objective 1: Minimize warpage deviation ,in, The number of sampling periods. For the first Warpage deviation per sampling period

[0225] Objective 2: Minimize the rate of change of PID parameter adjustments

[0226] Objective 3: Minimize the number of actuator adjustments (Defined as the number of times the actuator's movement exceeds the dead zone threshold)

[0227] Multi-objective optimization algorithm: Use NSGA-II (non-dominated sorting genetic algorithm) or MOPSO (multi-objective particle swarm optimization algorithm) to generate Pareto optimal solution set.

[0228] Training process:

[0229] Step 1: Data Preparation. Extract 1000 sets of typical operating condition data from the historical database as training samples. Each set of samples contains 5 process parameters, warpage deviation, and the corresponding optimal PID parameter adjustment. Divide the data into a training set (800 sets) and a validation set (200 sets).

[0230] Step 2: Encoding and Initialization. The neural network weight vectors are encoded as chromosomes, with the chromosome length equal to the total number of weights (taking DBN in Section 2.5 as an example, the total number of weights is approximately...). The initial population size is set to 100, using random initialization, with weights ranging from... .

[0231] Step 3: Fitness Calculation. For each individual in the population (i.e., a set of neural network weights), substitute it into the network, perform forward computation using the training set data, and obtain the fitness score for each sample. The average value is taken as the three objective function values ​​for that individual.

[0232] Step 4: Non-dominated ranking. The population is stratified according to Pareto dominance: if all goals of individual A are not inferior to those of individual B, and at least one goal of A is superior to that of B, then A dominates B. Individuals not dominated by any other individual constitute the first non-dominated layer (Pareto front), and so on to obtain all layers.

[0233] Step 5: Crowding Distance Calculation. Within the same non-dominated layer, calculate the crowding distance of each individual in the target space. The larger the distance, the sparser the surrounding area of ​​the individual, and the better the diversity.

[0234] Step 6: Selection. A tournament selection strategy is used, randomly selecting two individuals each time, prioritizing individuals with lower non-dominated layer numbers; if they are in the same layer, individuals with greater crowding distance are selected.

[0235] Step 7: Crossover and Mutation. Perform simulated binary crossover (SBX) on the selected individuals with a crossover probability of 0.9; perform polynomial mutation with a mutation probability of... Generate a progeny population.

[0236] Step 8: Elite Preservation. Merge the parent and offspring populations, recalculate the non-dominated ranking and crowding, and select the top 100 individuals as the new generation population.

[0237] Step 9: Iteration. Repeat steps 3-8 for a total of 500 iterations. Output the non-dominated solution set of the current Pareto front every 50 iterations.

[0238] Step 10: Pareto solution set generation. The final result is a set containing 20-30 Pareto optimal solutions, each corresponding to a set of neural network weights. The solutions in the set have varying degrees of effectiveness across the three objectives.

[0239] Online Selection Method (TOPSIS):

[0240] Before actual production begins, suitable weights are selected from the Pareto optimal solution set as initial weights based on the characteristics of the current production conditions. This embodiment uses TOPSIS (Topology-Optimal Solution Ranking) for multi-attribute decision-making, with the following specific steps:

[0241] 1. Construct the decision matrix. Given... There are 3 Pareto solutions, each with 3 objective values. Construct a matrix. ,in, .

[0242] 2. Normalization. Vector normalization is used: The normalized matrix is ​​obtained. .

[0243] 3. Determine the weights. Set target weights based on the current production conditions: If the current order has extremely high requirements for warpage (such as export packaging), set a weight vector. If high energy consumption and stability are required, then... Unless otherwise specified, assume The weights sum to 1.

[0244] 4. Construct a weighted decision matrix. .

[0245] 5. Determine the ideal solution and negative ideal solution For objective 1 ( Smaller is better), Target 2 ( Smaller is better), Target 3 ( (The smaller the better), positive ideal solutions take the minimum value of each column, and negative ideal solutions take the maximum value of each column:

[0246]

[0247]

[0248] 6. Calculate the Euclidean distance between each solution and the positive and negative ideal solutions:

[0249]

[0250]

[0251] 7. Calculate the relative closeness: , , The larger the value, the better the solution.

[0252] 8. Select the solution with the highest degree of approximation as the adaptive solution for the current working condition, and use its corresponding neural network weights as the initial weights.

[0253] If the current operating condition has a special preference for a certain objective, the weight vector can be adjusted. Recalculate and implement adaptive weight selection based on operating conditions.

[0254] Example 7: System Workflow and Performance Comparison

[0255] 1. System Workflow

[0256] This embodiment illustrates the complete workflow of the system.

[0257] Step 1: System Initialization

[0258] Load the weights of an offline trained artificial neural network model (backpropagation, deep neural network, or CNN can be selected).

[0259] Retrieve similar operating conditions from the historical database and initialize the baseline PID parameters. ;

[0260] The initial threshold is set according to national standards. (like , );

[0261] Step 2: Real-time data acquisition

[0262] The data acquisition unit collects five process parameters at a frequency of 10Hz.

[0263] The warpage detection device simultaneously detects the actual warpage value. ;

[0264] Calculate the deviation ;

[0265] Step 3: Forward computation of fuzzy neural network

[0266] The fuzzification module converts process parameters into fuzzy input vectors;

[0267] The neural network performs forward propagation and outputs... ;

[0268] Step 4: PID parameter self-tuning

[0269] according to The current PID parameters are calculated using the following formulas;

[0270] Step 5: PID Control and Feedforward Compensation

[0271] PID controller calculates control quantity ;

[0272] Feedforward compensation module calculation And by superimposing, we get ;

[0273] Step 6: Adjustment of the actuator

[0274] According to the execution unit Adjust the preheating cylinder heating valve, tension roller motor, and main drive motor;

[0275] Step 7: Hierarchical Trigger Judgment

[0276] The tiered trigger optimization module is based on Determine the interval to which it belongs:

[0277] like Keep the weights unchanged and return to step 2;

[0278] like Store the data in the historical database and return to step 2;

[0279] like : Triggers online fine-tuning, updating the neural network weights;

[0280] Step 8: Online Fine-tuning

[0281] Call the incremental learning algorithm, with a step size (BP network) or (DBN / CNN) Update weights;

[0282] continuous The triggering condition is met in all cycles before the event is truly triggered. (Set to 5)

[0283] The updated weights are used for forward calculation in the next cycle;

[0284] Step 9: Data Storage and Offline Optimization

[0285] Historical databases continuously store production data;

[0286] We perform offline retraining every weekend, using nearly three months of data to optimize the neural network model.

[0287] The dynamic threshold model and compensation coefficients are updated quarterly.

[0288] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Various changes made within the scope of knowledge possessed by those skilled in the art without departing from the concept of the present invention still fall within the scope of protection of the present invention.

Claims

1. A self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network, characterized in that, include: The data acquisition unit includes multiple sensors for real-time acquisition of process parameters during the corrugated cardboard production process. These process parameters include at least the moisture content of the base paper, the temperature of the preheating cylinder, the relative humidity of the environment, the speed of the production line, and the composite tension. The warpage detection device is installed at the end of the corrugated cardboard production line to detect the warpage direction and amount of the cardboard in real time and generate the actual warpage value. A fuzzy neural network controller is connected to both the data acquisition unit and the warpage detection device. It incorporates an offline-trained artificial neural network model and takes the process parameters as input, outputting adjustments to the PID control parameters, including a proportional gain adjustment. Integral coefficient adjustment and differential coefficient adjustment amount ; The PID parameter self-tuning unit is connected to the fuzzy neural network controller and is used to calculate the PID control parameters at the current moment in real time based on the reference PID parameters and the adjustment amount. A PID controller, connected to the PID parameter self-tuning unit, is used to calculate the control quantity based on the PID control parameters and warpage deviation value at the current moment. An actuator unit, connected to the PID controller, is used to receive the control quantity and adjust the operating parameters of the corrugated cardboard production line; The graded triggering optimization module is connected to both the warpage detection device and the fuzzy neural network controller. It is used to monitor the actual warpage value in real time and, based on the superior grade threshold defined in the corrugated cardboard product quality grade standard, optimize the warpage accordingly. and qualified product threshold Hierarchical trigger control strategy: When the actual warpage deviation satisfy At the same time, the current weights of the artificial neural network model remain unchanged; when At that time, the process parameters and deviation data at the current moment are stored in the historical database, but the online weight update of the artificial neural network model is not triggered. when When this occurs, the fuzzy neural network controller is triggered to perform online fine-tuning and update the weights of the artificial neural network model; Among them, the superior grade threshold and qualified product threshold It is pre-set according to the warpage limits for superior and qualified products specified in the national standards for the target corrugated cardboard products, and .

2. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, The process parameters are selected based on the multi-field coupling physical mechanism of corrugated cardboard warping deformation: The moisture content of the base paper and the relative humidity of the environment together constitute the factors affecting the moisture field, which are used to characterize the influence of the moisture migration trend of the paperboard on warping during the production process. The temperature of the preheating cylinder constitutes a factor affecting the temperature field and is used to characterize the effect of the degree of heat absorption of the cardboard in the preheating zone on warping. Production line speed constitutes a time-space coupled influencing factor, used to characterize the impact of the dwell time of the cardboard in each process section on moisture evaporation and heat transfer sufficiency; Composite tension constitutes a mechanical field influencing factor, used to characterize the effect of tensile stress on the stress distribution within the paperboard when the face paper and corrugated paper are bonded together.

3. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, The fuzzy neural network controller includes a fuzzification processing module and a neural network training module: The blurring processing module includes: Membership function definition unit, used to define multiple fuzzy subsets for each process parameter, and set membership functions for each fuzzy subset; The fuzzy rule base storage unit is used to store the fuzzy control rule base built based on the multi-field coupling mechanism; The fuzzy inference unit is used to perform fuzzy inference based on the continuous values ​​of each process parameter at the current time, combined with the membership function and the fuzzy control rules, to generate a fuzzy input vector. The fuzzy rule base includes at least: Rule R1: IF Preheat cylinder temperature = high AND paper moisture content = high AND machine speed = fast THEN Warping tendency = severe upward warping; Rule R2: IF Preheat cylinder temperature = low AND base paper moisture content = low AND machine speed = slow THEN Warp tendency = slight downward warping; Rule R3: IF Composite tension = High AND Preheat cylinder temperature = Normal AND Base paper moisture content = Normal THEN Warp tendency = Lateral warp; Rule R4: IF Ambient relative humidity = high AND base paper moisture content = normal AND preheating cylinder temperature = normal THEN Warping tendency = slight downward warping; Rule R5: IF Ambient relative humidity = low AND base paper moisture content = normal AND preheating cylinder temperature = normal THEN Warping tendency = slight upward warping; The neural network training module has the artificial neural network model built in, takes the fuzzy input vector as input, and outputs the adjustment amount of the PID control parameters.

4. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, In the hierarchical trigger optimization module, the trigger condition for online fine-tuning is: continuous Within each sampling period, the absolute value of the deviation between the actual warp value and the target warp value exceeded the qualified product threshold. ,in, Integers between 3 and 10; The online fine-tuning employs incremental learning algorithms, including stochastic gradient descent or mini-batch gradient descent, to update the weights of the artificial neural network model. The update step size is set to the initial learning rate during the offline training phase. to .

5. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, The artificial neural network model in the fuzzy neural network controller adopts a deep belief network or convolutional neural network structure to extract deep nonlinear coupling features between process parameters and warping deformation; the deep belief network or convolutional neural network is trained offline by combining unsupervised pre-training and supervised fine-tuning.

6. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, In the graded triggering optimization module, the superior product threshold and qualified product threshold It is a variable threshold that is dynamically adjusted based on warping data under the same or similar production conditions stored in a historical database, through cluster analysis or regression prediction models.

7. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, It also includes a baseline PID parameter adaptive initialization module, which is connected to the historical database. This module is used to retrieve the optimal baseline PID parameters under similar working conditions from the historical database before each production start, based on the cardboard specifications and raw material batch information of the current production order, and use these parameters as the initial baseline parameters. , , .

8. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, The graded triggering optimization module also includes a working condition similarity calculation unit, used to calculate the working condition similarity between the current production batch and historical batches stored in the historical database. The working condition similarity is calculated based on at least one factor among the raw paper batch, the rate of change of ambient temperature and humidity, and the equipment running time. When the working condition similarity exceeds a preset threshold, the historical threshold under the same or similar working conditions is preferentially used as the initial superior product threshold for the current batch. and qualified product threshold .

9. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, It also includes a control quantity feedforward compensation module, which is connected to the data acquisition unit and the PID controller. This module is used to pre-calculate the warping trend change based on the current rate of change in the raw paper moisture content and the rate of change in the production line speed, and generate a feedforward compensation amount that is superimposed on the control quantity output by the PID controller. The formula for calculating the feedforward compensation amount is: ,in, This refers to the moisture content of the base paper. For production line speed, and The compensation coefficient is determined based on historical data.

10. The self-tuning PID control system for corrugated cardboard warpage deformation based on a fuzzy neural network according to claim 1, characterized in that, The artificial neural network model in the fuzzy neural network controller is trained offline using a multi-objective optimization algorithm. The multi-objective includes at least two of the following: minimizing warpage deviation, minimizing the rate of change of PID parameter adjustment, and minimizing the number of actuator adjustments. The multi-objective optimization algorithm uses NSGA-II or MOPSO algorithm to generate a Pareto optimal solution set, and selects suitable neural network weights from the Pareto optimal solution set as initial weights according to the current production conditions.

Images